publications

We introduce TDiff-HSI, a diffusion-based model that can generate hyperspectral images (HSIs) directly from RGB images and material-wise segmentation masks. HSI provides both spatial (u, v) and spectral (λ) information. The accompanying dataset that we are releasing spans wavelengths in the range from 420 to 1728 nm, digitized into 512 channels. Directly handling this immense three-dimensional dataset is computationally prohibitive and often leads to numerical errors. To address this challenge, TDiff-HSI leverages Tucker decomposition to reduce dimensionality, enabling more stable and efficient processing. Moreover, spectral precision is enhanced by combining RGB channels with a material segmentation mask. To support this research, we constructed a new dataset using a hyperspectral camera. The dataset comprises 40,014 RGB-HSI pairs across 78 scenes, featuring 12 objects with diverse material properties.

We introduce BISeg, a novel hyperspectral image (HSI) segmentation model that achieves both accuracy and interpretability through band-interval selection. Unlike conventional approaches that rely on post-hoc band attribution, BISeg intrinsically learns to identify compact spectral windows during training. Each interval is parameterized by its center wavelength and width, with a maximum of ten intervals activated. A dedicated regularization encourages the use of fewer and narrower intervals, ensuring explanations remain concise and physically meaningful. The proposed spectral–spatial backbone integrates these intervals with spatial features via cross-axis attention, and a query-based head outputs pixel-level mask. Experiments on our custom HSI dataset, covering nine material classes, demonstrate that BISeg achieves competitive segmentation accuracy while producing interpretable spectral intervals aligned with known material properties.

We propose R2H-Diff, a diffusion-based model that reconstructs hyperspectral images (HSIs) using RGB images and semantic masks. Instead of generating the full spectrum at once, our model sequentially restores each spectral channel for greater training stability. A single-scene experiment validates its ability to denoise and reconstruct spectral data effectively.